Most secondary-data projects fail at the first decision rather than the last. Choosing among the public health databases for research determines what your study can possibly conclude, and no amount of careful analysis afterwards rescues a question pointed at the wrong source. This page sets out what each of the four major options answers, what each cannot answer whatever you do to it, and how to tell which one fits before you commit months of work.
In Simple Words
Each database was built to answer one kind of question. NHANES tells you how common something is across the country. SEER tells you how long cancer patients live. NTDB tells you what happens to injured patients in hospital. Genetic data tells you whether one thing actually causes another. Pick the wrong one and reviewers will reject the paper no matter how good the statistics are.
The Short Answer
| If your question is | Use | It cannot tell you |
|---|---|---|
| How common is X in the US population, or is X associated with Y across the general public | NHANES | Anything longitudinal, rare disease, hospital outcomes, treatment effects |
| How long do patients with this cancer survive, and how has that changed | SEER | Comorbidity, performance status, chemotherapy regimen, recurrence |
| What happens to injured patients across many trauma centres | NTDB / TQP | Population incidence or rates, anything after discharge |
| Does this exposure actually cause that outcome | Mendelian randomization | The effect of intervening at one point in life, or anything without strong instruments |
If none of those rows matches your question cleanly, the honest answer is usually that the question needs reshaping rather than a fifth database. That is what a feasibility check is for.
NHANES: How Common, and Associated With What
The National Health and Nutrition Examination Survey combines an interview with a physical examination and laboratory work on a stratified probability sample of the non-institutionalised US population. Because the sample is drawn to represent the country, correctly weighted estimates generalise nationally, which no hospital series can claim.
Its strength is breadth of measurement on ordinary people. Dietary intake, biomarkers, anthropometry, physical activity, environmental exposures and self-reported conditions all sit in the same file, so cross-domain questions are possible. Mortality linkage through the National Death Index adds a survival dimension.
Its limitation is structural and permanent. Every participant is measured once, so exposure and outcome are recorded at the same moment and temporality cannot be established. Describing a NHANES association in causal language is one of the fastest routes to rejection.
The design also has to be respected in the analysis or the numbers are simply wrong. We tested this directly and found that ignoring the survey weights inflated a national prevalence estimate by 45 percent, from 10.23 percent to 14.85 percent, on the same 5,018 people. The full test, with the code, is in what happened when we ran the same NHANES analysis three ways.
Access: fully public, no application, no agreement. Service: NHANES Study, $1,900, 15 business days.
SEER: Cancer Survival and Trends
The Surveillance, Epidemiology and End Results programme covers roughly half the US population through population-based cancer registries, recording incidence, tumour characteristics, stage, first-course treatment and cause of death with long follow-up. For uncommon tumours where no trial exists and no single institution has enough cases, it is often the only serious evidence available.
Use it for survival by stage and histology, incidence and outcome trends over time, and disparities by race, insurance or geography. Those are the questions it was designed to answer and it answers them well.
What it does not record is the difficulty. There is no comorbidity index, no performance status, no detailed regimen, no recurrence, and nothing about treatment beyond first course. Treatment comparisons drawn from SEER alone are confounded by indication in ways statistical adjustment does not fully remove, and oncology reviewers know it. Competing risks are the other trap, because in an older cancer cohort death from other causes is a competing event rather than censoring, and Kaplan-Meier applied to cancer-specific death in that setting overstates the event probability. The full guide to what the registry holds, which grouping to use now that SEER 22 has become SEER 21, and the five analytical errors that account for most rejections is in SEER database explained.
Access: either route, we hold access or work from your SEER*Stat extract. Service: SEER Cancer Outcomes Study, $2,100, 15 business days.
NTDB and TQP: Trauma Outcomes Across Centres
The National Trauma Data Bank and its successor, the American College of Surgeons Trauma Quality Programs participant use file, aggregate trauma registry records from participating centres. The scale allows questions about injury patterns, complications, disparities in trauma care and centre-level variation that no single hospital could support.
The critical constraint is that this is a convenience sample of participating centres, not a population-based registry. The denominator is unknown, the participating hospital set changes year to year, and patients who died in the field or attended non-participating centres are absent. Any paper reporting a national injury rate from these data is wrong, and trauma reviewers will say so. There is also no post-discharge follow-up, so long-term function and readmission are unavailable. The same caution about what public injury data can carry applies to related sources, as we discuss in whether NEISS injury statistics are reliable.
Analytically, injury severity has to be derived rather than trusted, physiologic variables are missing in patterns that correlate with severity, and patients cluster within centres. Registry data of this kind also needs disciplined cleaning, covered in six techniques for cleaning clinical registry data. The full guide to what NTDB and TQP data can and cannot support, including how injury severity is coded and how facility clustering should be modelled, is in NTDB and TQP trauma registry data explained. We built an independent simulation showing that undocumented complete-case handling of missing physiologic data roughly halves a cohort’s reported mortality rate, from 5.50% to 2.68%, detailed in how much complete-case analysis actually biases trauma mortality.
Access: you must hold the ACS data use agreement and supply the file. We cannot obtain it for you. Service: NTDB / TQP Trauma Registry Study, $2,400, 15 business days.
Mendelian Randomization: Does It Actually Cause It
The other three answer descriptive and prognostic questions. When the question is genuinely causal and a randomised trial is not feasible, Mendelian randomization uses genetic variants as instrumental variables, exploiting the fact that alleles are allocated at conception independently of most confounders.
It suits modifiable exposures with well-powered genome-wide association data, particularly lifelong exposure to a biomarker or a body composition measure, and it is increasingly used to triage drug targets before development. Background on the method is in our explainer on what Mendelian randomization is and when to use it, and on the wider family in instrumental variable approaches.
Two cautions matter. The estimate is a lifelong average effect, not the effect of intervening at one point in life, and reviewers now expect papers to say so. And because the software is free and the summary statistics are public, the field filled with single-exposure papers carrying no sensitivity analysis, which is exactly what editors now reject on sight. Instrument strength, Steiger filtering, several pleiotropy-robust estimators, outlier detection and replication are no longer optional extras. We simulated 135,000 MR studies to measure what instrument strength actually buys you, and found that at an F-statistic of 3 the confidence interval is 13.6 times wider than the effect being measured. The full test, with both engines, is in what the F-statistic actually tells you. On the pleiotropy side, our simulation of what each of the four estimators actually protects you from shows that agreement between them is not proof the instruments are valid.
Access: public GWAS summary statistics, we obtain them. Service: Mendelian Randomization Study, $2,900, 15 business days.
What All Four Have in Common
Whichever source you use, the same handful of things decide whether the paper survives review.
- Novelty established before analysis, not after. These datasets saturate quickly. Maupin and colleagues, writing in BMC Medicine in 2026, searched publications from 2021 to July 2025 and found substantial redundancy built on NHANES specifically, defining redundancy as the same exposures against the same outcomes in the same population. Reviewers arrive expecting a template.
- The design stated explicitly in the methods. Naming the weight, strata and cluster variables, or the competing-risks estimator, or the clustering unit, signals competence in a few lines. Their absence signals the opposite just as fast.
- Reporting standards followed. STROBE for observational work, STROBE-MR for genetic causal inference, TRIPOD for prediction models. Our STROBE tutorial covers what reviewers actually check.
- Registration considered. Increasingly expected even for retrospective secondary analysis, as set out in why a retrospective study should still be registered.
- Limitations that name the real threat. Cross-sectional temporality, confounding by indication, convenience sampling, lifelong-effect interpretation. A limitations paragraph that lists generic caveats while omitting the obvious one reads as evasion. For where each design sits on the hierarchy, see levels of evidence in clinical research.
Still Not Sure Which One
Three questions usually settle it. Does your question need a national denominator? That points to NHANES and rules out NTDB. Does it need time-to-event follow-up? That points to SEER and rules out NHANES. Does it need to establish that one thing causes another, rather than that they occur together? That points to Mendelian randomization and rules out all three of the others.
If your question survives all three and still fits nowhere, or if it spans more than one source, it is an extended engagement rather than a fixed package.
Tools and Resources
Free Editverse Tools
GraphPro builds publication-ready charts, 25 types, with built-in t-test, ANOVA, correlation, and chi-square p-values. Free, no signup.
StatPro offers free statistical calculators and a test-selector tool covering group comparisons, categorical outcomes, time-to-event analysis, and diagnostic accuracy.
Flowchart Pro generates PRISMA 2020, PRISMA-ScR, CONSORT, STROBE, STARD, MOOSE, and ARRIVE flow diagrams directly in your browser.
Professional Support
We build studies on all four sources to the standards described above, and we say no when the data cannot answer the question.
Feasibility & Protocol Assessment
$450
Start here if you are unsure which database fits. Checks whether the question is already answered, whether the variables exist, and whether you have the power. Credited in full against any study you commission.
$1,900
Nationally representative prevalence and exposure-outcome work, with the full complex survey design applied.
$2,100
Survival, incidence trends and disparities, with competing risks and proportional hazards handled properly.
$2,400
Risk-adjusted trauma outcomes with facility clustering modelled. You supply the data under your ACS agreement.
$2,900
Two-sample causal inference with the full robustness battery journals now require.
Quoted
Linked data, mediation designs, prediction models with external validation. Typically from $3,500.
The full category is at Registry and Real-World Data Studies.
References
- Maupin D, Suchak T, Barnett A, Spick M. Dramatic increases in redundant publications in the Generative AI era. BMC Medicine. 2026;24(1):29. doi:10.1186/s12916-025-04569-y
- National Center for Health Statistics. NHANES Survey Methods and Analytic Guidelines. CDC
- National Cancer Institute. Surveillance, Epidemiology, and End Results Program. seer.cancer.gov
- American College of Surgeons. Trauma Quality Programs Participant Use File. facs.org
- von Elm E, Altman DG, Egger M, et al. The STROBE statement. The Lancet. 2007;370(9596):1453-1457. doi:10.1016/S0140-6736(07)61602-X
- Skrivankova VW, Richmond RC, Woolf BAR, et al. Strengthening the reporting of observational studies in epidemiology using Mendelian randomization: the STROBE-MR statement. JAMA. 2021;326(16):1614-1621. doi:10.1001/jama.2021.18236